Presentation Creator — Guardrails

Phase 4 runs two complementary checkers against outline.yaml:

Treat the emitted pattern_history.history_source as the sole selected source. Treat recurring_antipatterns as the final filtered catalog recurrence contract. guardrail-check.py recurring_pattern_history_items() owns recurrence selection and optional movement attachment. Do not reproduce either decision in prose. pattern_history.suppressed_fields records each unavailable history-derived field. Current-outline contextual taxonomy scanning and illustration coverage (checks 9B and 10 below) remain agent-run checks alongside the script output.

The JSON object contains schema_version, talk_title, named checks, the full pattern_history authorization payload, structured recurring_antipatterns, contextual_taxonomy_scan, and required_companion_check. A produced report exits 0 even when a check status is FAIL. Input failures exit non-zero with diagnostics on stderr and no stdout.

The two scripts are independent — run both. check-rhetorical.py needs no profile and emits a deterministic report regardless of the speaker. guardrail-check.py fuses outline data with profile thresholds and venue context.

This file defines the guardrail check structure — what each check covers, how it's wired to schema fields, and how results are reported.

Thresholds and speaker-specific rules come from the vault at runtime:

• Slide budget tables → speaker-profile.json → guardrail_sources.slide_budgets[]
• Act 1 ratio limits → speaker-profile.json → guardrail_sources.act1_ratio_limits[]
• Recurring issues → speaker-profile.json → guardrail_sources.recurring_issues[]
• Branding checklist → speaker-profile.json → design_rules.footer
• Profanity rules → speaker-profile.json → rhetoric_defaults
• Confirmed intents → speaker-profile.json → confirmed_intents[]

If the speaker profile is not available, Section 16 remains usable for confirmed intent. Do not resolve Section 15 catalog history manually. Use the guardrail report's emitted history_source, available_classification_domains, and reasons. When the report selects no catalog-history source, use Section 15 only as narrative/audit context and run the current taxonomy scan without historical labels.

If policy_semantic_sha256 changes within one catalog/scoring generation, treat classification differences as a comparison reset, never as speaker improvement or regression.

Run these checks after Phase 3 delivery and after each Phase 4 revision.

1. Slide Budget

Read guardrail_sources.slide_budgets[] from the speaker profile for the thresholds. Match the talk's duration to the closest budget entry.

Progressive-reveal slides count toward the budget. If you show the same chart 3 times with different bars highlighted, that's 3 slides, not 1. Flag progressive reveals as budget-expensive and ask if the emphasis is worth the cost.

Demo-driven talks are the exception. If the selected mode is demo-driven, apply a much lower slide count (the live demo IS the content).

Check

[PASS/FAIL] Slide count: {actual}/{budget} for {duration}-minute slot

If over budget, suggest specific cuts. Prioritize cutting:

1. Progressive reveals that could be combined into one slide
2. Meme-only slides in Act 1 (meme accretion is a common pattern)
3. Redundant evidence slides (keep the strongest 2-3)

2. Act 1 Ratio (Problem Section Balance)

Read guardrail_sources.act1_ratio_limits[] from the speaker profile for the limits.

Check

[PASS/WARN/FAIL] Act 1 ratio: {act1_slides}/{total_slides} = {percentage}%
                 (limit: {max}% for {duration}-min slot)

WARN if within 5% of limit. FAIL if over.

3. Conference Branding Checklist

Read design_rules.footer from the speaker profile for the footer elements. Generate the checklist dynamically from footer.elements[]:

• Each footer element matches THIS venue
• Co-presenter element correct (if applicable, per footer.co_presented_extra)
• No stale conference names in slides
• No stale conference logos
• Shownotes URL slug matches this talk

Check

[PASS/FAIL] Branding: Footer elements specified for {conference}
[WARN] Branding: Conference hashtag not yet confirmed — flag for author

4. Profanity Audit

Read rhetoric_defaults.profanity_calibration and rhetoric_defaults.on_slide_profanity from the speaker profile.

The key rule (common across speakers): keep profanity verbal-only by default. On-slide profanity limits deck reuse across venues.

Check

[PASS/FAIL] Profanity register: {spec_register} applied consistently
[WARN/FAIL] On-slide profanity: {count} instances found — {approved/not approved}

If on-slide profanity is present without explicit approval, flag it: "Slide {N} has baked-in profanity: '{text}'. This limits reuse at corporate/ family-friendly events. Keep it verbal-only, or explicitly approve for this talk?"

5. Data Attribution

Survey data without visible source attribution creates credibility risk.

Check every data slide for:

• Source name visible
• Citation ID if available
• No orphaned percentages without context ("84% of developers..." — 84% of what survey?)

[PASS/FAIL] Data attribution: {N} data slides checked, {M} missing sources
            Slides needing sources: {list}

6. Time-Sensitive Content

Expired dates, deadlines, and promotional material appear on reused slides.

Check for:

• Registration deadlines
• Early-bird pricing
• "Coming soon" for things already launched
• Conference-specific dates
• Version numbers that may be outdated
• Cultural references that have expired (memes have a half-life)

[PASS/FAIL] Time-sensitive content: {count} items found
            {list with slide numbers and content}

7. Closing Completeness

Even compressed formats need a structured close. Read rhetoric_defaults.three_part_close from the speaker profile for whether the speaker defaults to a full three-part close.

Minimum viable close:

• Summary: At least 2-3 bullet points crystallizing the argument
• CTA: At least one specific, actionable next step
• Social: Speaker handles + shownotes URL

Check

[PASS/FAIL] Closing: {summary present?} | {CTA present?} | {social present?}
[WARN] Closing: Missing {component} — intentional or oversight?

8. Modular Cut Lines

Always check this. Read rhetoric_defaults.default_duration_minutes and rhetoric_defaults.modular_design from the speaker profile.

In outline.yaml, cuttable chapters and slides carry cuttable: true. The guardrail counts cuttable minutes and verifies the talk can compress to shorter slots.

Check

[PASS/FAIL] Cut lines: {cuttable_min} min of cuttable content
            (cuttable chapters: {ids}; cuttable slides: {ns})

FAIL when the profile's rhetoric_defaults.modular_design is true and no chapter or slide carries cuttable: true. PASS-and-skipped when modular_design is false (speakers who opt out don't get penalized for inflexible decks). The talk-duration-shorter-than-default heuristic is no longer used — the modular_design flag is the single gate.

9. Anti-Pattern Flags

This check has two layers: speaker-specific recurring issues from the vault, and taxonomy-based antipattern scanning from the Presentation Patterns reference.

9A. Speaker-Specific Recurring Issues

Read guardrail_sources.recurring_issues[] from the speaker profile. Each entry describes a known weakness and its specific guardrail check.

Schema-v4/v5 top-level entries are an independent lane: use an entry only when it carries exact source_lane: "non_pattern". Suppress a legacy or ambiguous entry instead of guessing its provenance. Catalog-derived recurring warnings live only in the guarded pattern_profile.antipattern_classifications lane below; do not duplicate them here. For each authorized non-pattern entry, run the check described in its guardrail field and report at the severity level in its severity field.

Common anti-patterns (may or may not apply to a given speaker):

• Meme accretion: If Act 1 has more than 60% meme/image-only slides, flag it
• Theoretical framing delay: If an opening framework section exceeds 10% of time, flag it
• Missing anti-sell beat: If commercial intent + applicable mode but no anti-sell
• Edgy humor overshoot: Flag culturally sensitive humor targets

9B. Presentation Patterns Taxonomy Scan

Read references/patterns/_index.md (Phase 4 section of the phase-grouped lookup table) and classification-derived profile history only when the Phase 0 pattern-history status lists the domain required for that claim.

Speaker-specific antipatterns. Emit one [RECURRING] line per recurring_antipatterns record in the guardrail report. Preserve each record's pattern_id, recurrence_classification, evidence, and optional trend. Do not re-open profile rows or re-filter the emitted records. An empty array produces no [RECURRING] line.

Contextual antipatterns — scan the outline against ALL antipatterns from the taxonomy. For each match, read the individual pattern file for detection heuristics and scoring criteria. These are flagged as [CONTEXTUAL] (a current-outline detection, not a historical claim). The current-outline scan always runs, including when pattern history is disabled.

Contextual detection rules:

• Bullet-Riddled Corpse — flag slides with 5+ bullet points or complete sentences
• Ant Fonts — flag text descriptions suggesting small or cramped content
• Cookie Cutter — flag ideas awkwardly split across slide boundaries
• Shortchanged — flag if duration is shorter than speaker's default with no cut lines
• Dual-Headed Monster — flag if co-presented with no handoff protocol
• Slideuments — flag if outline suggests dual-purpose slides (handout + presentation)
• Borrowed Shoes — flag if adapting another speaker's materials
• Dead Demo — flag demos without clear narrative purpose
• Alienating Artifact — flag humor/references that could exclude audience segments
• Tower of Babble — flag unexplained jargon for the stated audience level

Report format:

[RECURRING] {pattern_id} ({recurrence_classification} | {evidence} | {optional trend})
[CONTEXTUAL] Meme accretion — Act 1 meme ratio at 55%
[CONTEXTUAL] Bullet-Riddled Corpse — slides 14, 22 have 6+ bullet points
[CONTEXTUAL] Dual-Headed Monster — co-presented talk, handoff points not defined

When recurring_antipatterns is empty, omit recurring lines and keep contextual lines. If all history is disabled, also include the exact status warning. Do not relabel a contextual detection as recurring because Section 15 mentions a similar phrase.

10. Illustration Coverage

Only runs when the outline includes an Illustration Style Anchor section. When no illustration strategy exists, output [SKIP] Illustrations: no illustration strategy defined in the guardrail report and skip all sub-checks below.

Checks

Visual coverage ratio — what percentage of non-EXCEPTION slides have image prompts. Not every slide needs an illustration (text-only slides, transition slides), but the ratio should reflect the author's intent for the deck's visual density.

Format consistency — every slide has a Format field (FULL, IMG+TXT, EXCEPTION, or a custom format from the format vocabulary).

EXCEPTION justification — every EXCEPTION slide has a reason explaining why it uses a real asset instead of a generated illustration.

Style anchor reference — every image prompt starts with [STYLE ANCHOR] (the token the generation script replaces with the actual anchor text).

Prompt quality — no prompts that are just the Illustration description copy-pasted. The image prompt should be richer and more specific than the human-readable description (it includes composition details, labeling, specific visual elements).

Check

[PASS/FAIL] Illustration coverage: {N}/{M} illustrated slides have image prompts
[PASS/FAIL] Format tags: {N}/{total} slides have format specified
[WARN] EXCEPTION slides: {N} without justification
[PASS/FAIL] Style anchor reference: {N}/{M} prompts start with [STYLE ANCHOR]
[PASS/WARN] Prompt quality: {N} prompts appear to be copy-paste of description

Build coverage check — when slides define - Builds: sections:

[PASS/SKIP] Build coverage: {N} slides have builds defined, {M} build-step images generated

SKIP if no slides define builds. PASS if all defined build steps have corresponding images in illustrations/builds/. Report missing steps if any.

Prompt quality anti-patterns — flag prompts likely to produce poor results:

[WARN] Prompt anti-patterns: {N} prompts with potential issues

Detection rules (check all edit/fix/generation prompts):

• Missing preservation instructions: Flag edit prompts that remove or change content without explicit "keep [X]" instructions. Gemini removes elements it was not asked to remove — a prompt saying "remove the label" without "keep the soldiers, keep the border" will often damage unrelated elements.
• Missing safety suffixes: Flag edit prompts missing "DO NOT add any new elements" (Gemini aggressively adds unwanted decorative elements during edits) and/or "let background continue naturally" (Gemini fills erased areas with flat-colored patches).
• Simplified style anchor: Flag prompts that use a shortened or paraphrased version of the style anchor instead of the full text. The full specificity IS what produces the distinctive style — a prompt using "military style, pen and ink" instead of the full multi-line anchor will produce generic results. Heuristic: if a generation prompt's style description is significantly shorter than the anchor, flag it.
• Content addition via edit: Flag prompts that attempt to ADD new visual content (new characters, new objects, new text) using image editing. Content additions should use regeneration from the full prompt, not editing — editing strips the style when the model draws new elements, producing style mismatches.
• PIL/programmatic manipulation: Flag any use of PIL, Pillow, ImageMagick, or programmatic image manipulation for creating builds or fixing illustrations. These produce visible texture mismatches. Always use the model's native image editing.

11. Murder-Your-Darlings Filter Pass

Always runs. This is the convergent-thinking cut pass that asks of every section in the outline: does this directly support the Big Idea? Per the "Murder Your Darlings — The Pre-Delivery Cut Pass" subsection in patterns/prepare/crucible.md.

Check

For each top-level section in the outline:

• Identify which Big Idea component the section serves (POV reinforcement / stakes elaboration / proof / call-to-action setup / new-bliss preparation / inoculation).
• If you cannot name the component a section serves, the section is a candidate for the cut — flag it.
• Especially scrutinize sections marked [CALLBACK: master-story reference], sections labeled with a personal anecdote that doesn't tie back to the thesis, and sections containing data-rich material that took significant effort to gather. These are the most-loved-by-author and most-likely-to-be-darlings.
• The cut is rarely "remove the section entirely" — usually it's "compress to one beat" or "move into the speaker notes for reference."

[PASS/WARN/FAIL] Big Idea alignment: {N}/{total} sections traceable to Big Idea components
[WARN] Candidate darlings: {section names} — flagged for review

The pass produces flags, not automatic deletions. The author makes the actual cut decisions; the guardrail surfaces the candidates.

12. Emotion-Balance Check

Always runs. Per patterns/build/sparkline.md (the "Three Contrast Types" subsection) and Duarte's kairos principle, the analytical/emotional ratio of the outline must match the audience type.

Check

Categorize each content section as analytical (data, logic, argument, technical detail) or emotional (story, image, anecdote, evocative language). Compare the ratio against the audience profile from the spec:

• Analytical audience (scientists, engineers, accountants, board members): aim for 70–80% analytical content, 20–30% emotional content. Less emotion than this strips persuasive power; more reads as manipulation.
• Broad audience (general professional, mixed-discipline conference): aim for 50–60% analytical, 40–50% emotional.
• Emotionally-charged context (advocacy, fundraising, cause-driven keynote): can run 30–40% analytical, 60–70% emotional. Not less analytical than 30% even here — emotion without proof loses credibility.

[PASS/WARN] Emotion balance: {N}% analytical / {M}% emotional vs. {target}% target for {audience type}

WARN if the ratio is more than 15 percentage points off the target in either direction. The check is descriptive, not prescriptive — the author decides whether the imbalance is intentional (some talks deliberately skew). Surface the imbalance so it's a conscious choice.

13. Screening with Critics — Pre-Lock Gate

Runs before final spec lock for high-stakes talks (keynotes, sales pitches, fundraising presentations, executive briefings). Per the "Screening with Critics — Beyond Copyediting" subsection in patterns/build/peer-review.md.

Check

Has the author scheduled (or executed) a formal screening session with the following properties?

• 3× the duration of the presentation (60-min talk → 3-hour screening)
• Screeners selected from outside the speaker's organization (or, if internal, explicitly free of the six dysfunctional review patterns: Conceited Captain / Political Paranoia / Message Magic / Vacuum Visionary / Lackey Leader / Customer Cold-Shoulder)
• Screeners match the target audience profile (industry, role, knowledge level)
• Each screener has access to slides + speaker notes for line-level critique
• Output: structural critiques + language critiques + resistance discoveries (which feed back into Phase 3 inoculation moves)

[PASS/WARN/SKIP] Screening: {scheduled / completed / not applicable for this stakes level}

SKIP for low-stakes talks (internal demos, small-group presentations, tutorial sessions). For everything else, WARN if no screening is scheduled or completed.

Guardrail Summary Template

Render this human summary from the JSON report after completing the agent-run checks. This template is not the script's stdout contract.

GUARDRAIL CHECK — {talk title} — {date}
================================================
[PASS/WARN] Pattern history: {available domains / disabled with exact reasons}
[PASS] Slide budget: {actual}/{max} for {duration}-min slot
[PASS/WARN/FAIL] Act 1 ratio: {%} (limit: {max}%)
[PASS/FAIL] Branding: {status}
[PASS/FAIL] Profanity: {register} applied, {on-slide count} on-slide
[PASS/FAIL] Data attribution: {N} slides checked, {M} issues
[PASS/FAIL] Time-sensitive: {count} items
[PASS/FAIL] Closing: summary={y/n} CTA={y/n} social={y/n}
[PASS/FAIL] Cut lines: {present/missing}
[INFO/SKIP] Historical anti-patterns: {authorized recurring issues or suppressed}
[RECURRING] Presentation Patterns: {recurring_antipatterns records | omit when empty}
[CONTEXTUAL] Presentation Patterns: {current-outline taxonomy flags; always enabled}
[PASS/FAIL/SKIP] Illustrations: {coverage ratio} | {format tags} | {prompt quality}
[PASS/SKIP] Builds: {N} defined, {M} images generated
[WARN] Prompt anti-patterns: {N} issues found
[PASS/WARN/FAIL] Big Idea alignment: {N}/{total} sections traceable to Big Idea
[PASS/WARN] Emotion balance: {N}% analytical / {M}% emotional vs. {target}%
[PASS/WARN/SKIP] Screening with critics: {scheduled / completed / not applicable}
================================================

The Illustrations, Builds, and Prompt anti-patterns lines show [SKIP] when no illustration strategy is defined.